MLLMs Hallucinate when Information Distribution Drifts in Synergy Heads

类型：arxiv
标识：2609.09206
链接：https://arxiv.org/abs/2609.09206
主分类：multimodal
形态：application
TLDR：Multimodal Large Language Models (MLLMs) often struggle with hallucinations, thus hindering their reliable practical applications. Existing attention-based mitigation methods mainly rely on indirect signals (e.g., attention weights) that fail to accurately reflect the actual information shift underlying hallucination generation. In this paper, we propose HEAL, Head-lEvel information disentAnglement and caLibration for identifying and mitigating hallucinations. HEAL first employs causal noise intervention on multi-head outputs to filter out causally redundant heads. Subsequently, it disentangle
待LLM分类：否
来源文件：
/inbox/tom/_candidates/2026-09-21-agent-rag-longcontext-candidates.json

2026-08-30

笔记 Cool Papers · cs.CL (papers.cool) · RSS 摘要

2026-09-13

笔记 2026-09-18 flyP 精读：M3Exam 多模态记忆基准 + 2026 Agent Eval Stack

2026-09-18

笔记 2026-07-04 上午多模态对照精读 · Seeker（文本→像素的长上下文 MLLM）

2026-07-04

笔记 xHC: Expanded Hyper-Connections · 精读与批判(2026-07-21 09:50)

2026-07-21

文档 Stephen 自测 · R62 · 2026-09-01

2026-09-01

arXiv

2609.09206

TLDR

Multimodal Large Language Models (MLLMs) often struggle with hallucinations, thus hindering their reliable practical applications. Existing attention-based mitigation methods mainly rely on indirect signals (e.g., attention weights) that fail to accurately reflect the actual information shift underlying hallucination generation. In this paper, we propose HEAL, Head-lEvel information disentAnglement and caLibration for identifying and mitigating hallucinations. HEAL first employs causal noise intervention on multi-head outputs to filter out causally redundant heads. Subsequently, it disentangle

MLLMs Hallucinate when Information Distribution Drifts in Synergy Heads

相关 · 按语义匹配